Some investors, analysts, and industry watchers believe the big AI labs, especially Anthropic, are taking advantage of the current AI safety scare by inviting regulations that only the largest labs can easily comply with. The AI safety concerns are real, but the financial burden of complying with the proposed regulatory solutions would conveniently entrench the big labs, the thinking goes, while putting smaller labs and open-model developers at a serious disadvantage.
In this light, the recent safety-related announcements from the big labs, and their executives’ calls for slowing AI model development, can look performative, like AI safety “theater.” “Anthropic and Open AI are really good at flashing something in front of you and making it a big thing, but there’s always a strategy behind it,” says PitchBook senior analyst Harrison Rolfes.
Taken to its logical conclusion, critics argue, this kind of “regulatory capture” could reduce the number of tech companies providing the world’s AI to OpenAI, Anthropic, SpaceXAI, Google, Microsoft, and perhaps Mistral in Europe.
The “regulator” in this case might not be the federal or state government. Rather, the big labs could agree on a set of independent evaluation firms, perhaps organizations like Model Evaluation and Threat Research (METR), Redwood Research, and Apollo Research, that would test the safety and alignment of new frontier models. Rolfes says it might look something like the Big Four accounting firms, whose auditors are paid to independently vet companies’ financial statements.
In an essay published earlier this month, Anthropic CEO Dario Amodei proposed that frontier AI companies should have independent safety evaluators embedded inside them, with employee-like access, rather than leaving companies as the sole judges of their own safety. Anthropic’s policy page also says AI companies “shouldn’t be the only ones deciding whether their systems are safe” and calls for independent third-party evaluations. OpenAI CEO Sam Altman subsequently tweeted his agreement with Amodei’s proposal: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”
Rolfes says such a mechanism could create a kind of “moat” around the largest AI labs, since smaller labs and new market entrants could have a harder time shouldering the costs of undergoing these evaluations. “It actually adds a very large cost because now you need a third-party stamp of approval to release models,” Rolfes says. Independent evaluators would likely charge hefty fees, he argues, given the amount of work and expertise needed to clear a huge and complex model for public release.
Independent evaluators may also need substantial computing resources to run increasingly complex tests. So the costs grow with the complexity of the models. “Since models are now capable enough of acting as LM agents, evals have to be increasingly complex tasks, which significantly increases the overhead per eval,” wrote Marius Hobbhahn, CEO of the independent evaluation company Apollo Research, back in 2024.
Anthropic itself has acknowledged the risk and explicitly connected it to regulatory capture. “The costs of evaluating huge and complex AI models is high, and getting higher,” the company stated in a 2024 blog. “We should ensure this is a well scoped, small set of tests, or else we’ll create regulatory burdens and increase the possibility of regulatory capture.”
The heads of the big labs don’t often agree on much, but lately they’ve been singing from the same hymnal when it comes to AI safety and alignment. In a September 12 essay, Anthropic’s Amodei called for slowing, or “pacing,” frontier model development so that the labs could catch up on safety and alignment R&D. OpenAI’s Sam Altman publicly agreed, saying “we need to pace the frontier.” SpaceXAI’s Elon Musk simply responded, “Dario is right,” aligning himself with the argument for slowing down. Google DeepMind founder Demis Hassabis also expressed support.
Days later, OpenAI policy chief Chris Lehane told reporters that OpenAI had already been working with Anthropic and Google DeepMind on AI safety for several weeks.
Wired’s Maxwell Zeff reported on September 10 that OpenAI had asked members of Congress whether orchestrating an industry-wide slowdown in frontier AI development would violate antitrust laws, citing people close to the company. But the reporting doesn’t say OpenAI asked about an agreement among industry players on independent evaluations specifically. (OpenAI declined to comment for this story. Anthropic did not respond to a request for comment.)
It’s also possible that the labs are motivated by genuine safety concerns, or by some mix of motives. “This is a complex topic since it is hard to fathom the real motives,” says Gartner analyst Arun Chandrasekaran in an email to Fast Company. “[A]nthropic’s safety advocacy did contribute to its models later being restricted from government use,” he recalls. He’s referring to Anthropic’s limits on the Pentagon’s use of its models for domestic mass surveillance, an application the company said it had no reliable way to monitor for safety.
“I believe sincere risk concern shown by frontier labs and competitive advantage aren’t mutually exclusive,” Chandrasekaran says. “They may gain competitive advantage, but they also realize how much safety and alignment is critical [to] how their customers will evaluate them in the present and future.”
On Wednesday, OpenAI reported six more incidents from the past six months in which it discovered “unexpected or concerning” model behavior during training or evaluation. AI researchers call this misalignment—when models act outside their intended role, evade oversight, take unauthorized actions, or coordinate in unexpected ways. In July, swarms of agents powered by OpenAI models broke out of their testing environment, accessed the open internet, and broke into servers belonging to the open-source AI repository Hugging Face.
Anthropic has repeatedly disclosed that its own models attempted harmful actions under testing. In August, U.K. AI Security Institute testers observed Mythos 5 taking 17 unauthorized actions against third parties, including trying to plant malicious code in an open-source project and creating fake identities for a social-engineering attack. In a September 9 report, Anthropic described four occasions during 2026 when its Claude model accessed the public internet because of a “misconfiguration” in a cybersecurity evaluation designed by a testing partner.
David Sacks, a science and technology adviser to President Trump, has embraced the regulatory-capture theory to describe Anthropic’s statements and policy actions. “Dario’s post assumes we have amnesia about Anthropic’s well-orchestrated campaigns hyping AI fears,” he tweeted in response to Amodei’s September 12 essay. “His May 2025 claim that AI would wipe out 50% of entry-level knowledge jobs within five years still lacks supporting evidence 15 months later.” Sacks has explicitly accused Anthropic of pursuing a “sophisticated regulatory capture strategy,” arguing that its push for stringent AI-safety laws could impose compliance costs that smaller competitors can’t afford.
Why would these big, well-financed labs need to go to such lengths to freeze out competition through regulation or independent auditors? Being a freestanding frontier model developer—meaning one not housed inside a larger tech company—is a fabulously expensive business. The companies spend massive amounts renting or building cloud infrastructure to run their models, and they must reserve computing power years into the future.
A July report from The Information says Anthropic had previously told investors it had committed to spending $180 billion on computing resources through 2029, but that the company had signed commitments worth an estimated $517 billion for compute-capacity leases in the past 11 months.
OpenAI has said it will spend $750 billion on infrastructure through 2030, or 25% more than it estimated earlier in the year. CEO Sam Altman said last year that OpenAI had committed to spending $1.4 trillion to develop 30 gigawatts of computing resources through 2032.
The labs also have to pay the salaries of the world’s best AI researchers. The Wall Street Journal reported that OpenAI’s average stock-based compensation is about $1.5 million per employee across a workforce of about 4,000. Add model-training costs and other expenses, and the companies have enormous bills to cover.
Their revenues, while enormous and rapidly growing, remain much smaller than some of those long-term commitments. OpenAI’s annualized revenue run rate reached $40 billion in August, Bloomberg reported. Anthropic’s annualized revenue run rate passed $65 billion at the end of July, Bloomberg reported.
“So, Anthropic has to make so much money just to pay off all these commitments, and there’s just no way they’re going to be able to make that much unless they find a way to capture the entire market,” Rolfes says. He argues that Anthropic, OpenAI, and SpaceXAI each target different market segments, with some overlap, so if regulatory burdens rendered smaller labs uncompetitive, the remaining players could gain something approaching a monopoly or duopoly in portions of their target markets.
Rolfes’ team just released a new research report examining Anthropic’s business quality, revenue outlook, and compute strategy ahead of its S-1 filing, with the aim of evaluating whether its reported $2 trillion valuation is justified.
Safety may be the goal, but it could also become a barrier to entry.